Magnetic Target Detection Using PointRend-Based Region-Convolutional Neural Network

Abstract

The improving performance of high-resolution magnetic instruments has propelled the refined detection of small targets of different shapes. The traditional method is to extract target-related features through some empirical formulas, and then classify the targets visually. Such manually designed formulas have limited feature expression ability. Also, the results interpreted by different people may vary. Although machine learning was later applied to the task, which avoided the link of manual judgment, the accuracy and robustness were not strong. In this paper, an end-to-end Region Convolutional Neural Network (R-CNN) is proposed to identify targets with little human intervention. Considering the differences between magnetic signals and natural images in terms of scenario, target and imaging, improvements to R-CNN meta-architecture are required. Specifically, magnetic tensor gradient (MTG) data with grid cells are transformed into 2-D matrix, which is then enhanced by pseudo-color coding for mapping. Given the self-built dataset, we design a Two-stage Fine-grained (TSFG) R-CNN to excavate effective deep-level features of targets. PointRend is used here to predict the high-quality edge segmentation of targets. Experiment results show that the proposed method provides a useful way for the detection of multi-scale, multi-shape and multi-depth magnetic targets, even under the case of magnetic field superimposition.